University of Kentucky University of Kentucky

UKnowledge UKnowledge

Sanders-Brown Center on Aging Faculty Publications Aging

2-6-2020

Hierarchical Clustering Analyses of Plasma Proteins in Subjects Hierarchical Clustering Analyses of Plasma Proteins in Subjects

With Cardiovascular Risk Factors Identify Informative Subsets With Cardiovascular Risk Factors Identify Informative Subsets

Based on Differential Levels of Angiogenic and Inflammatory Based on Differential Levels of Angiogenic and Inflammatory

Biomarkers Biomarkers

Zachary Winder University of Kentucky, winder.zachary@uky.edu

Tiffany L. Sudduth University of Kentucky, tlsudd2@uky.edu

David W. Fardo University of Kentucky, david.fardo@uky.edu

Qiang Cheng University of Kentucky, Qiang.Cheng@uky.edu

Larry B. Goldstein University of Kentucky, larry.goldstein@uky.edu

See next page for additional authors

Follow this and additional works at: https://uknowledge.uky.edu/sbcoa_facpub

Part of the Biostatistics Commons, Computer Sciences Commons, Geriatrics Commons, Medical

Physiology Commons, Neurology Commons, and the Pathology Commons

Right click to open a feedback form in a new tab to let us know how this document benefits you. Right click to open a feedback form in a new tab to let us know how this document benefits you.

Repository Citation Repository Citation Winder, Zachary; Sudduth, Tiffany L.; Fardo, David W.; Cheng, Qiang; Goldstein, Larry B.; Nelson, Peter T.; Schmitt, Frederick A.; Jicha, Gregory A.; and Wilcock, Donna M., "Hierarchical Clustering Analyses of Plasma Proteins in Subjects With Cardiovascular Risk Factors Identify Informative Subsets Based on Differential Levels of Angiogenic and Inflammatory Biomarkers" (2020). Sanders-Brown Center on Aging Faculty Publications. 136. https://uknowledge.uky.edu/sbcoa_facpub/136

This Article is brought to you for free and open access by the Aging at UKnowledge. It has been accepted for inclusion in Sanders-Brown Center on Aging Faculty Publications by an authorized administrator of UKnowledge. For more information, please contact UKnowledge@lsv.uky.edu.

Hierarchical Clustering Analyses of Plasma Proteins in Subjects With Hierarchical Clustering Analyses of Plasma Proteins in Subjects With Cardiovascular Risk Factors Identify Informative Subsets Based on Differential Cardiovascular Risk Factors Identify Informative Subsets Based on Differential Levels of Angiogenic and Inflammatory Biomarkers Levels of Angiogenic and Inflammatory Biomarkers

Digital Object Identifier (DOI) https://doi.org/10.3389/fnins.2020.00084

Notes/Citation Information Notes/Citation Information Published in Frontiers in Neuroscience, v. 14, article 84, p. 1-10.

© 2020 Winder, Sudduth, Fardo, Cheng, Goldstein, Nelson, Schmitt, Jicha and Wilcock.

This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

Authors Authors Zachary Winder, Tiffany L. Sudduth, David W. Fardo, Qiang Cheng, Larry B. Goldstein, Peter T. Nelson, Frederick A. Schmitt, Gregory A. Jicha, and Donna M. Wilcock

This article is available at UKnowledge: https://uknowledge.uky.edu/sbcoa_facpub/136

fnins-14-00084 February 4, 2020 Time: 17:12 # 1

ORIGINAL RESEARCHpublished: 06 February 2020

doi: 10.3389/fnins.2020.00084

Edited by:Emmanuel Pinteaux,

The University of Manchester,United Kingdom

Reviewed by:Hilario Blasco-Fontecilla,

Puerta de Hierro University Hospital,Spain

Fanny M. Elahi,University of California,

San Francisco, United States

*Correspondence:Donna M. Wilcock

donna.wilcock@uky.edu

Specialty section:This article was submitted to

Neurodegeneration,a section of the journal

Frontiers in Neuroscience

Received: 15 May 2019Accepted: 21 January 2020

Published: 06 February 2020

Citation:Winder Z, Sudduth TL, Fardo D,

Cheng Q, Goldstein LB, Nelson PT,Schmitt FA, Jicha GA and Wilcock DM

(2020) Hierarchical ClusteringAnalyses of Plasma Proteins

in Subjects With Cardiovascular RiskFactors Identify Informative Subsets

Based on Differential Levelsof Angiogenic and Inflammatory

Biomarkers. Front. Neurosci. 14:84.doi: 10.3389/fnins.2020.00084

Hierarchical Clustering Analyses ofPlasma Proteins in Subjects WithCardiovascular Risk Factors IdentifyInformative Subsets Based onDifferential Levels of Angiogenic andInflammatory BiomarkersZachary Winder1,2, Tiffany L. Sudduth1, David Fardo1,3, Qiang Cheng4,Larry B. Goldstein5, Peter T. Nelson1,6, Frederick A. Schmitt1,5, Gregory A. Jicha1,5 andDonna M. Wilcock1,2*1 Sanders-Brown Center on Aging, University of Kentucky, Lexington, KY, United States, 2 Department of Physiology,University of Kentucky, Lexington, KY, United States, 3 Department of Biostatistics, University of Kentucky, Lexington, KY,United States, 4 Department of Computer Science, University of Kentucky, Lexington, KY, United States, 5 Department ofNeurology, University of Kentucky, Lexington, KY, United States, 6 Department of Pathology, University of Kentucky,Lexington, KY, United States

Agglomerative hierarchical clustering analysis (HCA) is a commonly used unsupervisedmachine learning approach for identifying informative natural clusters of observations.HCA is performed by calculating a pairwise dissimilarity matrix and then clustering similarobservations until all observations are grouped within a cluster. Verifying the empiricalclusters produced by HCA is complex and not well studied in biomedical applications.Here, we demonstrate the comparability of a novel HCA technique with one that wasused in previous biomedical applications while applying both techniques to plasmaangiogenic (FGF, FLT, PIGF, Tie-2, VEGF, VEGF-D) and inflammatory (MMP1, MMP3,MMP9, IL8, TNFα) protein data to identify informative subsets of individuals. Studysubjects were diagnosed with mild cognitive impairment due to cerebrovascular disease(MCI-CVD). Through comparison of the two HCA techniques, we were able to identifysubsets of individuals, based on differences in VEGF (p < 0.001), MMP1 (p < 0.001),and IL8 (p < 0.001) levels. These profiles provide novel insights into angiogenic andinflammatory pathologies that may contribute to VCID.

Keywords: hierarchical clustering analysis, vascular cognitive impairment and dementia, mild cognitiveimpairment, VEGF, MMP1, IL8

INTRODUCTION

Vascular cognitive impairment and dementia (VCID) is an active area in dementia research(Murphy et al., 2016) and is described as “encompassing all the cognitive disorders associated withcerebrovascular disease (CVD), from dementia to mild cognitive deficits” (Gorelick et al., 2011).VCID is estimated to occur in roughly 20% of the cases of dementia; however, the exact prevalencein the population is unknown with varying estimates in the literature (Rizzi et al., 2014; Corriveauet al., 2016). Much of the uncertainty in assessing the prevalence of VCID is due to varied diagnosticcriteria (Harrison et al., 2016). In addition, there is substantial overlap in cognitive manifestationsof cerebrovascular and neurodegenerative pathologies [such as Alzheimer’s disease (AD)] thatcan culminate in clinical dementia (Kapasi and Schneider, 2016), which further complicates our

Frontiers in Neuroscience | www.frontiersin.org 1 February 2020 | Volume 14 | Article 84

fnins-14-00084 February 4, 2020 Time: 17:12 # 2

Winder et al. Biomarkers of Small Vessel Disease

understanding of VCID. Further, both pathologies commonly co-exist in the same individual, yet some autopsy studies suggest thatthere is a significant increase in dementia risk due to vascularfactors when Alzheimer pathology is low (Abner et al., 2017;Dodge et al., 2017).

Currently, magnetic resonance imaging (MRI) andcerebrospinal fluid (CSF) biomarkers are used to differentiateVCID from AD and monitor the progression of VCID (Cipolliniet al., 2019; Wardlaw et al., 2019). Plasma biomarkers arecurrently being investigated as a lower cost and less invasivealternative approach. The current study is focused on exploringthe potential clustering of plasma biomarkers using hierarchicalclustering analysis (HCA) in participants with VCID who havemild cognitive impairment (MCI) due to CVD (MCI-CVD) toidentify unique plasma profiles of disease (Winblad et al., 2004).Persons with MCI-CVD are of particular interest as they are at anincreased risk of developing dementia and already have cognitivedecline (Petersen et al., 1999). We evaluated angiogenic (FGF,FLT, PIGF, Tie-2, VEGF, VEGF-D) and inflammatory (MMP1,MMP3, MMP9, IL8, TNFα) protein plasma biomarkers in theseparticipants using the highly sensitive meso-scale discoveries(MSD) platform. Angiogenic and inflammatory markers are ofparticular interest due to their roles in endothelial dysfunction,which has been shown to play a role in the pathogenesis of CVD(Poggesi et al., 2016; Mahoney et al., 2019). Presently, studieshave demonstrated mixed results in the association of angiogenicand inflammatory biomarkers with VCID; however, it issuspected that this is due to the inconsistency in both the patientpopulations and the analytical measures (Cipollini et al., 2019).

Agglomerative HCA is an unsupervised machine learningtechnique commonly used to determine similar subsets within alarger population (Xu and Wunsch, 2010). HCA can be used toidentify subsets within a variety of different patient populations.The accuracy of this technique is difficult to quantify, as moststudies rely on post hoc analysis of the clusters produced by HCAto determine their validity. We propose a unique methodologyfor validating clusters produced by HCA. This method relies onusing two unique HCA models on the same dataset and evaluatescongruencies between the two models by comparing a novel HCAmodel to one that is widely used (Wallin et al., 2010; Damianet al., 2013; Nettiksimmons et al., 2014; Racine et al., 2016). Beforeapplying both HCA models to our dataset, we tested the accuracyof each model on various distributions of data and comparedthem to each other using the adjusted rand index (ARI). Afterdemonstrating the interchangeability of the two HCA models inthe simulated data distribution comparable to our dataset, wetested both models on our dataset and compared the underlyingcomponents of each cluster produced by the HCA models.

MATERIALS AND METHODS

ParticipantsPlasma samples were collected from a cohort of adult researchvolunteers enrolled in a randomized behavioral interventionstudy for MCI-CVD (N = 80, NCT01924312). Inclusioncriteria for the parent study include age older than 55 years,

Montreal Cognitive Assessment score < 29, and at leastone uncontrolled vascular risk factor. Risk factors includedpoorly controlled hypertension, poorly controlled cholesterol,cardiomyopathy/CHF, diabetes with a fasting glucose > 110 orHbA1c > 7%, homocysteine > 12, history of transient ischemicattack, tobacco use > 30 pack-years, and BMI > 30. Potentialsubjects were excluded from this cohort if they had dementia,evidence of a non-CVD cause of cognitive decline, evidence of anon-CVD neurologic disease, or any focal motor, sensory, visual,or auditory deficits. For the current study, participants were alsoexcluded if they had an incomplete panel of markers as measuredvia MSD assays as described below (n = 7).

Plasma Collection and QuantificationPlasma samples were collected by venous puncture using10 mL EDTA Vacutainer tubes. Plasma was aliquot into cryo-tubes at 500 mL volumes. Quantification of plasma sampleswas accomplished using MSD Multi-Spot V-PLEX assays[Angiogenesis Panel 1 (human) and Proinflammatory Panel 1(human)] and Ultra-Sensitive assays (MMP 2-Plex and MMP 3-Plex). Plasma did not undergo any freeze-thaw cycles after theinitial thawing of the aliquot. Assays were performed using platespecific protocols as followed with analysis performed in the MSDDiscovery Workbench 4.0 software.

MMP 2-Plex and MMP 3-PlexMMP plates were brought to room temperature forapproximately 30 min and then loaded with 25 µL of diluent,covered (protect from light), and incubated at room temperaturefor 30 min while shaking at 600 r/min. After incubation, plateswere removed from the shaker and 25 µL of calibrator was addedto the assigned wells in duplicate with 5 µL of undiluted sampleand 20 µL of diluent. Plates were covered and incubated at roomtemperature while shaking at 600 r/min. After incubation, plateswere removed from the shaker and washed three times with300 µL of wash buffer. Plates were turned upside down andtapped against paper towels to ensure the removal of all washbuffer from the wells. 25 µL of the antibody mix was loaded intoeach well, covered (protect from light), and incubated at roomtemperature for 2 h shaking at 600 r/min. After incubation, plateswere removed from the shaker and the wash steps were repeatedfrom above; 150 µL of read buffer was loaded into each well andread on the MSD Quickplex SQ 120 machine.

Proinflammatory Panel 1Proinflammatory plates were brought to room temperature forapproximately 30 min and washed three times with 300 µLof wash buffer. Plates were turned upside down and tappedagainst paper towels to ensure the removal of all wash bufferfrom the wells; 50 µL of calibrator was added to the assignedwells in duplicate with 50 µL of undiluted sample and covered(protect from light). Plates were incubated at 4◦C overnight whileshaking at 600 r/min. In the morning, plates were removed from4◦C and incubated at room temperature for 1 h while shakingat 600 r/min. After incubation, plated were removed from theshaker and the wash steps were repeated from above; 25 µL ofthe antibody mix was added into each well, covered (protect

Frontiers in Neuroscience | www.frontiersin.org 2 February 2020 | Volume 14 | Article 84

fnins-14-00084 February 4, 2020 Time: 17:12 # 3

Winder et al. Biomarkers of Small Vessel Disease

from light), and incubated at room temperature for 2 h shakingat 600 r/min. After incubation, plates were removed from theshaker and wash steps were repeated from above; 150 µL ofread buffer was loaded into each well and read on the MSDQuickplex SQ 120 machine.

Angiogenesis Panel 1Angiogenesis plates were brought to room temperature forapproximately 30 min and then loaded with 150 µL of diluent,covered (protect from light), and incubated at room temperaturefor 1 h while shaking at 600 r/min. After incubation, plates wereremoved from the shaker and washed three times with 300 µL ofwash buffer. Plates were turned upside down and tapped againstpaper towels to ensure the removal of all wash buffer from thewells; 50 µL of calibrator was added to the assigned wells induplicate with 25 µL of undiluted sample and 25 µL of diluent.Plates were covered (protect from light) and incubated at 4◦Covernight while shaking at 600 r/min. In the morning, plates wereremoved from 4◦C and incubated at room temperature for onehour while shaking at 600 r/min. After incubation, plates wereremoved from the shaker and wash steps were repeated fromabove; 25 µL of the antibody mix was added into each well,covered (protect from light), and incubated at room temperaturefor 2 h shaking at 600 r/min. After incubation, plates wereremoved from the shaker and wash steps were repeated fromabove; 150 µL of read buffer was loaded into each well and readon the MSD Quickplex SQ 120 machine.

Samples were run in duplicate and three pooled controlsamples were run on each plate to measure inter- and intra-plate variability. MSD quantification was performed on a tablestabilizer in order to reduce error in MSD plate readings.

Plasma Sample AnalysisProtein markers measured through MSD assays were subjectedto intra- and inter-plate variability tests. Intra-plate variabilitywas assessed through two distinct methods. The first methodcalculated the percentage of samples for each marker that had acoefficient of variation, as determined by the duplicate runs foreach sample, greater than or equal to 0.25. Markers that contained20% of samples above this threshold were removed from furtheranalysis. The second method ran three pooled control sampletwice on the same plate (two samples each run in duplicate)to ensure consistency in final quantifications. The coefficientof variation for each of the three controls measured for eachmarker was then averaged together. Markers with an averagecoefficient of variation greater than 0.25 were excluded fromthe analysis. Markers that passed both criteria were included inthe final analysis. Inter-plate variability was accounted for usingthe three pooled control samples run on each plate. Each platecontrol value was divided by the control mean and all three ofthese values for each marker were averaged together to providea plate-scaling factor. Each value was then divided by this factorto adjust for inter-plate variability. The resulting measures werelog-transformed to scale each marker to a common order ofmagnitude, which is required in clustering algorithms to provideequal weighting of markers. Grubb’s test was lastly applied tothe data to remove outliers (Grubbs, 1950). Individual samples

containing one or more outliers in the measured markers wereexcluded from further analysis (n = 7) due to their effects onclustering techniques. The final dataset consisted of 66 patientplasma samples, which were quantified for 11 plasma markers(FGF, FLT, PLGF, Tie-2, VEGF, VEGFD, MMP1, MMP3, MMP9,IL8, TNFα).

Hierarchical Clustering AnalysisAll HCAs were performed using the Matlab Statistics andMachine Learning Toolbox functions pdist, linkage, and cluster.Previously described HCA models were comprised of threedifferent algorithms, distance, linkage, and clustering (Wallinet al., 2010; Damian et al., 2013; Nettiksimmons et al., 2014;Racine et al., 2016). The conventional HCA model consists ofa Euclidean distance algorithm, which calculates the distancebetween two samples using the Euclidean distance formula (aspecial case of the generalized Minkowski distance formula),where the distance between observations s and t in a sample withn markers equals dst :

dst =

√√√√ n∑j=1

∣∣Xsj − Xtj∣∣2

The linkage algorithm used was Ward’s Linkage, whichcalculates the incremental increase in within-cluster sum ofsquares and links samples one at a time until all samples arecombined into a single cluster (Xu and Wunsch, 2010). Thismethod combines similar samples until all samples fall withinone cluster (i.e., agglomerative hierarchical clustering). The finalalgorithm in the conventional HCA model used a standardagglomerative clustering approach (Racine et al., 2016).

The novel proposed HCA model uses consensus clusteringas presented by Fred and Jain (2002) to combine HCA modelswith different distance and clustering algorithms. The distancealgorithms used the Minkowski distance formula with p rangingfrom 0.1 to 2.0 in increments of 0.1. The distance betweenobservations s and t in a sample with n markers equals dst :

dst = p

√√√√ n∑j=1

∣∣Xsj − Xtj∣∣p

Each distance algorithm’s data were then used with theweighted average linkage algorithm, which combines samples

TABLE 1 | Means ± standard deviation for age, MMSE, and MoCA for theMCI-CVD cohort population in addition to percent of female participants.

Mean ± SDev Range

Age (years) 75.07 ± 8.14 (56.99–89.22)

MMSE 26.86 ± 2.95 (18–30)

MoCA 22.11 ± 3.74 (11–28)

Systolic blood pressure (mmHg) 141.33 ± 15.31 (102–185)

Hemoglobin A1c (%) 6.18 ± 1.31 (4.3–11.8)

LDL cholesterol (mg/dL) 97.44 ± 42.63 (22–299)

Sex 47% Female

Frontiers in Neuroscience | www.frontiersin.org 3 February 2020 | Volume 14 | Article 84

fnins-14-00084 February 4, 2020 Time: 17:12 # 4

Winder et al. Biomarkers of Small Vessel Disease

into clusters that have the smallest distance between them anddetermines that distance using a recursive function which treatsthe subset of linkages equally (Xu and Wunsch, 2010).

Lastly, data from each linkage algorithm were clusteredusing an inconsistency clustering algorithm. This algorithmcalculates an inconsistency coefficient of a new linkage usingthe mean and standard deviation of the linkage heights for aspecified depth (dep) of sub linkages within each new linkage.Clusters were formed when the inconsistency coefficient for eachlinkage and all sub linkages were less than a specified cutoff(cut) value. Each linkage algorithm output was run throughmultiple iterations of the inconsistency clustering algorithmwith values for depth (dep) from 2 to 6 in increments of1, whereas cutoff (cut) values were adjusted from 1.0 to 3.0in increments of 0.1. All iterations of depth and cutoff wereevaluated, and if only one cluster was formed, then thatiteration was not used in the consensus clustering model.Once each clustering model was established, distances betweensamples were calculated based on the percentage of modelsin which two samples shared a cluster. Samples that sharedno clusters were given a distance of 1 and, samples thatwere paired in the same cluster in each model were given adistance of 0. Plots of each clustering model were created usingthe dimensionality reduction function, t-distributed stochasticneighbor embedding (t-sne), with a random number generationseed of 10 to maintain reproducibility (Maaten, 2008). Clinicaldata were excluded from the clustering algorithm to avoidclusters based on clinical findings as this study sought to

identify clusters of participants based on a differential level offluid biomarkers.

Simulated Data Generation and AnalysisSimulated data generation was performed using the MatlabStatistics and Machine Learning Toolbox function mvnrnd. Eachsimulated data experiment was run with 35 trials and eachtrial was initiated with a unique random number generationseed to maintain reproducibility. Generated data contained 11variables and 100 samples per group, obtained from knowndistributions with the mean and sigma of each distributiondiffering depending on the experiment. Supplementary TablesS1–S3 detail the mean and sigma for each group within eachexperiment. The ARI was used to evaluate the accuracy ofeach clustering model by comparing each clustering result tothe known cluster assignment. The ARI has a maximum valueof 1 indicating that the clustering result matches perfectly tothe known cluster assignment. An ARI of 0 indicates that theclustering model assigns observations to the correct clusterassignment with an equal probability as random chance. An ARIbelow 0 demonstrates that the clustering model is less effectivethan random chance at assigning observations to the correctcluster assignment (McComb, 2015).

Statistical AnalysisStatistical analysis was performed using the Matlab Statistics andMachine Learning Toolbox and SPSS. A two-sample t-test usingthe Matlab function ttest2 was conducted to compare the ARI

FIGURE 1 | Compares the combined HCA model to the Euclidean distance model in eight different datasets. Each dataset was produced using the mvnrnd. Theaccuracy of each model was assessed using the adjusted rand index (ARI). Datasets used in A–D have identical means for each variable in each cluster. Datasetsused in E–H have the same means for the first six variables and means of opposite signs for the other five variables. Means and standard deviations for each datasetare shown in Supplementary Table S1. Red horizontal lines indicate means for each model. Stars (*) indicate statistical significance between groups using anindependent samples t-test (p < 0.05).

Frontiers in Neuroscience | www.frontiersin.org 4 February 2020 | Volume 14 | Article 84

fnins-14-00084 February 4, 2020 Time: 17:12 # 5

Winder et al. Biomarkers of Small Vessel Disease

FIGURE 2 | Compares the combined HCA model to the Euclidean distance model in two predicted distributions. Each dataset was produced using mvnrnd. Theaccuracy of each model was assessed using the adjusted rand index (ARI). Means and covariance matrix for each dataset are shown in Supplementary Tables S2,S3. Red horizontal lines indicate means for each model. Stars (*) indicate statistical significance between groups using an independent samples t-test (p < 0.05).(A) Shows estimated two cluster model, and (B) shows estimated three cluster model.

means of the two HCA models for all simulated data experiments.SPSS was used for the remaining statistical tests to determinedifferences between clusters for each log-transformed proteinmarker. Levene’s test for equality of variances was performedbefore each two-sample t-test, and Satterthwaite’s t-test was usedfor any marker found to have significantly different variances.Levene’s test for homogeneity of variances based on the meanwas also conducted before performing an ANOVA test foreach marker and a Welch’s test for equality of means wasperformed for markers with non-homogeneous variances. Posthoc analysis was then conducted on markers which had asignificant p-value for an ANOVA or Welch’s test. Tukey’s HSDwas used for significant ANOVA tests and Dunnett T3 was usedfor significant Welch’s test.

RESULTS

Study Population DescriptionDemographic and neurocognitive evaluations were obtained in65/66 participants within our MCI-CVD cohort (Table 1). Themean age of the participants was 75.07 ± 8.14 with a femalepopulation of 47%. MMSE scores ranged from 18 to 30 with amean of 26.86 ± 2.95, while MoCA scores ranged from 11 to28 with a mean of 22.11 ± 3.74. Vascular risk factors including

systolic blood pressure, hemoglobin A1C, and LDL cholesterolwere also evaluated in our cohort (Table 1). Mean systolic bloodpressure was found to be 141.33± 15.31 mmHg, hemoglobin A1c6.18± 1.31%, and LDL cholesterol 97.44± 42.63 mg/dL.

Simulated Data AnalysisTo test the applicability of the novel combined HCA model,we tested its accuracy in eight unique simulated datasets(detailed in Supplementary Table S1). We tested the novelmodel against an established HCA model using the ARI tomeasure the accuracy of each model (Figure 1). In our firstexperiment, we studied the accuracy of both models in a datasetwith two distant uniform clusters (Figure 1A). The establishedHCA model using Euclidean distance showed no difference inARI compared to the novel combined HCA model (Euclidean:0.9892 ± 0.0029, Novel: 0.9920 ± 0.0019, p = 0.413). A similarresult was found in a dataset with two distant variable clusters(Euclidean: 0.9920 ± 0.0023, Novel: 0.9926 ± 0.0020, p = 0.857)(Figure 1E). These results demonstrate that both models wereable to assign each distribution to its own cluster. Next, we testedboth models on a dataset with three distant uniform clusters(Figure 1C) and three distant variable clusters (Figure 1G).The established HCA model had a significantly increased ARIover the novel HCA model in both of these experiments(Euclidean: 0.6186 ± 0.0139, Novel: 0.4067 ± 0.0320, p < 0.001,

Frontiers in Neuroscience | www.frontiersin.org 5 February 2020 | Volume 14 | Article 84

fnins-14-00084 February 4, 2020 Time: 17:12 # 6

Winder et al. Biomarkers of Small Vessel Disease

FIGURE 3 | T-SNE dimensionality reduction plot of clustered data. Data were clustered using the Euclidean distance model (A) and the Combined HCA model (B).The dataset was produced by measuring 11 inflammatory (MMP1, MMP3, MMP9, IL8, TNFα) and angiogenic proteins of interest (FGF, FLT, PIGF, Tie-2, VEGF,VEGF-D) from 66 participant plasma samples.

FIGURE 4 | Compares clusters 1 and 2 produced from the Combined HCA model in each inflammatory and angiogenic protein measured. (A–K) Show mean andscatter of FGF (A), FLt (B), PlGF (C), Tie-2 (D), VEGF (E), VEGF-D (F), MMP1 (G), MMP3 (H), MM9 (I), IL8 (J), and TNFa (K). Red horizontal lines indicate themeans for each cluster. Stars (*) indicate statistical significance between groups as calculated by the log transform of the data shown using an independent samplest-test (p < 0.05). Means ± SEM are shown in Table 2.

Frontiers in Neuroscience | www.frontiersin.org 6 February 2020 | Volume 14 | Article 84

fnins-14-00084 February 4, 2020 Time: 17:12 # 7

Winder et al. Biomarkers of Small Vessel Disease

TABLE 2 | Means ± SEM for clusters 1 and 2 produced by the combined HCAmodel.

Biomarker p-value Cluster 1 (pg/mL) Cluster 2 (pg/mL)

FGF <0.001 24.56 ± 2.39 12.27 ± 1.22

FLT 0.568 12.92 ± 1.35 11.87 ± 0.59

PIGF 0.093 3.52 ± 0.38 2.94 ± 0.11

Tie-2 0.013 500.55 ± 31.69 411.09 ± 14.77

VEGF <0.001 98.39 ± 17.46 43.55 ± 3.63

VEGFD 0.350 560.73 ± 52.17 515.59 ± 24.35

MMP1 <0.001 5244.74 ± 700.11 2292.86 ± 219.55

MMP3 0.005 17275.46 ± 3267.49 10547.71 ± 856.39

MMP9 <0.001 43561.18 ± 8177.19 19232.98 ± 1851.63

IL8 <0.001 4.20 ± 0.49 2.25 ± 0.11

TNFa 0.283 1.81 ± 0.15 1.63 ± 0.10

Statistical significance between groups was determined using the log transform ofthe data shown in Figure 4 in an independent samples t-test (p < 0.05).

Figure 1C) (Euclidean: 0.5767± 0.0200, Novel: 0.4393± 0.0092,p < 0.001, Figure 1G). These results show that the establishedHCA model has a higher accuracy when separating threedistant clusters of normally distributed data. We then testedif the models performed differently on distributions thathad more overlapping characteristics. The first experimentsof these distributions were with two close uniform clusters(Figure 1B) and two close variable clusters (Figure 1F). In bothexperiments, the established HCA model had a higher accuracycompared to the novel HCA model (Euclidean: 0.6579 ± 0.0150,Novel: 0.5482 ± 0.0237, p < 0.001, Figure 1B) (Euclidean:0.6186 ± 0.0139, Novel: 0.4067, p < 0.001, Figure 1F). Thisdifference continued in the final set of experiments which usedthree close uniform clusters (Euclidean: 0.2477 ± 0.0087, Novel:0.2103± 0.0103, p< 0.007, Figure 1D) and three distant variableclusters (Euclidean: 0.2370 ± 0.0080, Novel: 0.2023 ± 0.0078,p < 0.003, Figure 1H). These experiments show that as thedistributions progressively overlap the accuracy for both modelsdecrease and the difference between the accuracy of the modelsdecreases as well.

Predicted Distribution AnalysisWe hypothesized that clusters, if any, in our empirical datasetwould overlap more and thus be more difficult to differentiatethan those used in the previous experiments. To test the accuracyof each model in this distribution, we generated simulateddata from predicted distributions based on analysis from ourcollaborators (detailed in Supplementary Tables S2, S3). Thefirst experiment was based on a two-cluster model within oursample population (Figure 2A). This experiment showed nodifferences between the established Euclidean HCA model andthe novel HCA model (Euclidean: 0.1422 ± 0.0118, Novel:0.1270 ± 0.0181, p = 0.486, Figure 2A). In addition, we testeda three cluster model for our sample population and foundsimilar results with no differences between the two models inthis study (Euclidean: 0.0902 ± 0.0092, Novel: 0.0895 ± 0.0105,p = 0.962, Figure 2B). The data from these two experimentsdemonstrate the interchangeability of these two models whenstudying datasets with extensive overlapping distributions.

Application of Models to DatasetAfter validating the novel combined HCA model using predicteddistributions, we applied both HCA models to our 66 patientsample (Figure 3). When the Euclidean distance HCA modelwas applied to our dataset, four clusters emerged (Figure 3A).Clusters 1, 3, and 4 appear to be more compact in the 2-D t-SNEdimensionality reduction plot, while cluster 2 exists along theperiphery of the plot in a more scattered manner. We continuedthis experiment and applied the novel combined HCA modelto the same dataset and uncovered two clusters (Figure 3B).Cluster 1 contains 14 samples of which 12 also appear incluster 1 of the Euclidean distance HCA model. The other 52samples appear in cluster 2, which is comprised of clusters2–4 from the Euclidean distance HCA model. The similarityof these two results emphasizes the underlying distributionswithin this dataset.

Characterizing Cluster DifferencesWe proceeded to analyze the differences that drive clusterdifferentiation. First, we examined the clusters produced by

TABLE 3 | Means ± SEM for clusters 1–4 produced by the Euclidean distance model.

Biomarker p-value Cluster 1 (pg/mL) Cluster 2 (pg/mL) Cluster 3 (pg/mL) Cluster 4 (pg/mL)

FGF <0.001 24.90 ± 2.78 5.16 ± 0.60 12.70 ± 1.37 20.11 ± 1.97

FLT 0.340 12.77 ± 1.56 13.08 ± 0.82 11.05 ± 1.44 11.42 ± 0.82

PIGF 0.342 3.53 ± 0.45 2.94 ± 0.17 2.79 ± 0.28 3.10 ± 0.16

Tie-2 0.040 512.42 ± 35.98 437.13 ± 25.44 384.29 ± 26.72 405.64 ± 21.72

VEGF <0.001 106.97 ± 19.29 49.97 ± 6.65 27.69 ± 4.39 48.56 ± 5.14

VEGFD 0.258 563.88 ± 60.58 579.18 ± 46.39 469.25 ± 36.56 487.25 ± 32.33

MMP1 <0.001 5605.20 ± 769.82 3061.39 ± 356.41 736.27 ± 71.84 2692.85 ± 285.62

MMP3 0.096 17517.09 ± 3822.29 10666.69 ± 1443.84 9646.52 ± 1192.19 11587.35 ± 1561.28

MMP9 <0.001 46055.29 ± 9378.94 25868.67 ± 3771.38 17474.83 ± 2787.13 14764.34 ± 1748.47

IL8 <0.001 4.49 ± 0.52 2.32 ± 0.18 1.78 ± 0.12 2.53 ± 0.18

TNFα 0.362 1.86 ± 0.16 1.52 ± 0.16 1.51 ± 0.10 1.83 ± 0.21

Statistical significance between groups was determined using the log transform of the data shown in Figure 5 in an ANOVA followed by Tukey’s HSD for post hoc analysis(p < 0.05). Results of the ANOVA omnibus are shown in p-value.

Frontiers in Neuroscience | www.frontiersin.org 7 February 2020 | Volume 14 | Article 84

fnins-14-00084 February 4, 2020 Time: 17:12 # 8

Winder et al. Biomarkers of Small Vessel Disease

FIGURE 5 | Compares clusters 1–4 produced from the Euclidean distance model in each inflammatory and angiogenic protein measured. (A–K) Show mean andscatter of FGF (A), FLt (B), PlGF (C), Tie-2 (D), VEGF (E), VEGF-D (F), MMP1 (G), MMP3 (H), MM9 (I), IL8 (J), and TNFa (K). Red horizontal lines indicate themeans for each cluster. Stars (*) indicate statistical significance between groups as calculated by the log transform of the data shown using ANOVA followed byTukey’s HSD for post hoc analysis (p < 0.05). Means ± SEM are shown in Table 3.

FIGURE 6 | Cluster comparison of VEGF (pg/mL), MMP1 (pg/mL), and IL8 (pg/mL) for data clustered using the Euclidean distance model (A) and the combinedHCA model (B) from 66 participant plasma samples.

Frontiers in Neuroscience | www.frontiersin.org 8 February 2020 | Volume 14 | Article 84

fnins-14-00084 February 4, 2020 Time: 17:12 # 9

Winder et al. Biomarkers of Small Vessel Disease

the novel HCA model (Figure 4), and found that cluster 1was increased compared to cluster 2 in FGF (p < 0.001),Tie-2 (p = 0.013), VEGF (p < 0.001), MMP1 (p < 0.001),MMP3 (p = 0.005), MMP9 (p < 0.001), and IL8 (p < 0.001)(Table 3). When the clusters produced from the Euclideandistance HCA model were analyzed (Figure 5), we found asimilar pattern of clusters. In this model, cluster 1 was increasedcompared to clusters 2–4 in VEGF (p = 0.006, p < 0.001,p = 0.013), MMP1 (p = 0.003, p < 0.001, p = 0.001), andIL8 (p < 0.001, p < 0.001, p < 0.001), respectively. Cluster 1was also increased in FGF (p = 0.004), Tie-2 (p = 0.033), andMMP9 (p = 0.003) compared to cluster 3. The elevated levelof these proteins in cluster 1 agrees with the characteristics ofcluster 1 established previously with both models demonstratinga subset with significant increases in VEGF, MMP1, andIL8 compared to the other subsets (Figure 6). However, theEuclidean distance HCA model does show differences betweenclusters 2 and 4, which were not seen in the novel HCAmodel. Clusters 2 and 4 were similar in their makeup, bothincreased over cluster 3 in MMP1 (p < 0.001 and p < 0.001)and VEGF (p = 0.032 and p = 0.016), respectively, butdifferent levels of FGF (p < 0.001). These differences lead tothe possibility of four disease profiles within the MCI-CVDpatient population.

DISCUSSION

The results of this study provide evidence supporting theuse of the novel combined HCA model in datasets withextensive overlapping distributions. The results of the firstset of experiments demonstrate that the Euclidean distanceHCA model outperforms the novel combined HCA model indatasets with a moderate amount of overlapping distributions(Figures 1B–D,F–H). This difference is reduced as thedistributions progressively increase in overlap and eventuallydisappears entirely in our second set of experiments involving themore complex datasets with predicted distributions (Figure 2).

It is important to note that the capacity of the twoHCA models to accurately cluster data into their knowndistributions decreases as the datasets become more complex.In our experiments involving two distant distributions of data,both models were able to separate each distribution with anARI approximately equal to 1 (Figures 1A,E). In experimentsusing our predicted distributions, the average ARI decreasedto approximately 0.13 and 0.09 for the two and three clustermodels, respectively (Figure 2). These findings demonstratethe limits of reliability in both HCA models and providea measure to compare additional HCA models to in futureexperiments. Accounting for this accuracy is crucial wheninterpreting HCA results because clusters produced by the HCAmodel may not correspond to any true unique distributionand may simply be a subset within the normal variation ofa larger distribution. Therefore, it is important to comparemultiple HCA models on an unknown dataset in order toelucidate which clusters are in fact unique distributions withinthe dataset. Overall, results from our experiments support the

interchangeability of HCA models in datasets similar to thoseshown in Supplementary Tables S2, S3, which allows forthe use of both models in assessing clustering distributionswithin our dataset.

Both the Euclidean distance HCA model and the novelcombined HCA model resulted in similar disease profileswithin our cohort of MCI-CVD patients (Figure 3). In thisstudy, both models classified participants into a cluster thathad elevated levels of VEGF, MMP1, and IL8 comparedto the other clusters (Figures 4, 5). We suspect that thisdisease profile seen in cluster 1 may be involved in a moreactive VCID process resulting in increased pathology due tothe increased level of angiogenic and inflammatory markers.Clusters 2–4 in the Euclidean distance HCA model mayalso be clinically relevant in terms of disease pathology butrequire future studies to understand how these profiles maycontribute to progression of VCID in a population of individualswith MCI-CVD.

CONCLUSION

The usage of both the novel HCA model and a Euclidean distanceHCA model identified a novel subset of patients within the MCI-CVD population. This study provides insight into a potentialunderlying inflammatory and angiogenic profile of disease inpatients with VCID. Defining subsets of patients within thispopulation with different disease profiles continues to be a keyresearch objective. These profiles can provide a more completeunderstanding of disease progression and allow physicians andresearchers to identify patients undergoing different rates ofpathologic change in a prospective cohort. In the future, we hopeto further clarify these profiles by combining plasma and MRIimaging biomarkers that can also be used in clinical trials as keyoutcome measures to determine the efficacy of novel therapeutics.

DATA AVAILABILITY STATEMENT

The datasets generated for this study are available on request tothe corresponding author.

ETHICS STATEMENT

The studies involving human participants were reviewed andapproved by the University of Kentucky Institutional ReviewBoard. The patients/participants provided their written informedconsent to participate in this study.

AUTHOR CONTRIBUTIONS

ZW designed and executed the analyses. TS performed theassessment of analytes using MSD. DF and QC consulted onstudy design. LG consulted on clinical readouts. PN providedsamples. FS consulted on neuropsychological assessments. GJ

Frontiers in Neuroscience | www.frontiersin.org 9 February 2020 | Volume 14 | Article 84

fnins-14-00084 February 4, 2020 Time: 17:12 # 10

Winder et al. Biomarkers of Small Vessel Disease

assessed the patients and collected all samples. DW oversawthe studies and assisted in the development of hypotheses anddata interpretation.

ACKNOWLEDGMENTS

The authors gratefully acknowledge the Sanders-Brown Centeron Aging Clinic team for their support in participant recruitmentand evaluations, Dr. Erin Abner for her assistance in study design,

the NIH [NINR: 4R01NR014189-05 (GJ), NIA: 5UH2NS100606-02 (DW and GJ), NCATS: UL1TR001998, NIA: 5P30AG028383],and the participants of this study for their time and commitment.

SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be foundonline at: https://www.frontiersin.org/articles/10.3389/fnins.2020.00084/full#supplementary-material

REFERENCESAbner, E. L., Kryscio, R. J., Schmitt, F. A., Fardo, D. W., Moga, D. C., Ighodaro,

E. T., et al. (2017). Outcomes after diagnosis of mild cognitive impairment in alarge autopsy series. Ann. Neurol. 81, 549–559. doi: 10.1002/ana.24903

Cipollini, V., Troili, F., and Giubilei, F. (2019). Emerging biomarkers in vascularcognitive impairment and dementia: from pathophysiological pathways toclinical application. Int. J. Mol. Sci. 20:2812 doi: 10.3390/ijms20112812

Corriveau, R. A., Bosetti, F., Emr, M., Gladman, J. T., Koenig, J. I., Moy, C. S.,et al. (2016). The science of vascular contributions to cognitive impairmentand dementia (VCID): a framework for advancing research priorities in thecerebrovascular biology of cognitive decline. Cell Mol. Neurobiol. 36, 281–288.doi: 10.1007/s10571-016-0334-7

Damian, M., Hausner, L., Jekel, K., Richter, M., Froelich, L., Almkvist, O., et al.(2013). Single-domain amnestic mild cognitive impairment identified by clusteranalysis predicts Alzheimer’s disease in the european prospective DESCRIPAstudy. Dement. Geriatr. Cogn. Disord. 36, 1–19. doi: 10.1159/000348354

Dodge, H. H., Zhu, J., Woltjer, R., Nelson, P. T., Bennett, D. A., Cairns, N. J., et al.(2017). Risk of incident clinical diagnosis of Alzheimer’s disease-type dementiaattributable to pathology-confirmed vascular disease. Alzheimers Dement. 13,613–623. doi: 10.1016/j.jalz.2016.11.003

Fred, A., and Jain, A. K. (2002). “Data clustering using evidence accumulation,”in Proceedings of the International Conference Pattern Recognition, Quebec,276–280.

Gorelick, P. B., Scuteri, A., Black, S. E., Decarli, C., Greenberg, S. M.,Iadecola, C., et al. (2011). Vascular contributions to cognitive impairment anddementia: a statement for healthcare professionals from the american heartassociation/american stroke association. Stroke 42, 2672–2713. doi: 10.1161/str.0b013e3182299496

Grubbs, F. E. (1950). Sample criteria for testing outlying observations. Ann. Math.Statist. 21, 27–58. doi: 10.1214/aoms/1177729885

Harrison, S. L., Tang, E. Y., Keage, H. A., Taylor, J. P., Allan, L., Robinson, L., et al.(2016). A systematic review of the definitions of vascular cognitive impairment.No dementia in cohort studies. Dement. Geriatr. Cogn. Disord. 42, 69–79. doi:10.1159/000448213

Kapasi, A., and Schneider, J. A. (2016). Vascular contributions to cognitiveimpairment, clinical Alzheimer’s disease, and dementia in older persons.Biochim. Biophys. Acta 1862, 878–886. doi: 10.1016/j.bbadis.2015.12.023

Maaten, L. (2008). Visualizing data using t-SNE. J. Mach. Learn. Res. 9:2579.Mahoney, E. R., Dumitrescu, L., Moore, A. M., Cambronero, F. E., De Jager, P. L.,

Koran, M. E. I., et al. (2019). Brain expression of the vascular endothelial growthfactor gene family in cognitive aging and alzheimer’s disease. Mol. Psychiatry[Epub ahead of print]. doi: 10.1038/s41380-019-0458-5

McComb, C. (2015). Adjusted RAND INDEX, 1.0.0.0 Edn. Natick, MA: MATLABCentral File Exchange.

Murphy, M. P., Corriveau, R. A., and Wilcock, D. M. (2016). Vascular contributionsto cognitive impairment and dementia (VCID). Biochim. Biophys. Acta 1862,857–859. doi: 10.1016/j.bbadis.2016.02.010

Nettiksimmons, J., Decarli, C., Landau, S., Beckett, L., and Alzheimer’s DiseaseNeuroimaging Initiative, (2014). Biological heterogeneity in ADNI amnesticmild cognitive impairment. Alzheimers Dement. 10:511-521.e1. doi: 10.1016/j.jalz.2013.09.003

Petersen, R. C., Smith, G. E., Waring, S. C., Ivnik, R. J., Tangalos, E. G., andKokmen, E. (1999). Mild cognitive impairment: clinical characterization andoutcome. Arch. Neurol. 56, 303–308.

Poggesi, A., Pasi, M., Pescini, F., Pantoni, L., and Inzitari, D. (2016). Circulatingbiologic markers of endothelial dysfunction in cerebral small vessel disease: areview. J. Cereb. Blood Flow Metab. 36, 72–94. doi: 10.1038/jcbfm.2015.116

Racine, A. M., Koscik, R. L., Berman, S. E., Nicholas, C. R., Clark, L. R.,Okonkwo, O. C., et al. (2016). Biomarker clusters are differentially associatedwith longitudinal cognitive decline in late midlife. Brain 139, 2261–2274. doi:10.1093/brain/aww142

Rizzi, L., Rosset, I., and Roriz-Cruz, M. (2014). Global epidemiology of dementia:Alzheimer’s and vascular types. Biomed. Res. Int. 2014:908915. doi: 10.1155/2014/908915

Wallin, A. K., Blennow, K., Zetterberg, H., Londos, E., Minthon, L., and Hansson,O. (2010). CSF biomarkers predict a more malignant outcome in Alzheimerdisease. Neurology 74, 1531–1537. doi: 10.1212/WNL.0b013e3181dd4dd8

Wardlaw, J. M., Smith, C., and Dichgans, M. (2019). Small vessel disease:mechanisms and clinical implications. Lancet Neurol. 18, 684–696. doi: 10.1016/S1474-4422(19)30079-1

Winblad, B., Palmer, K., Kivipelto, M., Jelic, V., Fratiglioni, L., Wahlund, L. O., et al.(2004). Mild cognitive impairment–beyond controversies, towards a consensus:report of the International working group on mild cognitive impairment.J. Intern. Med. 256, 240–246.

Xu, R., and Wunsch, D. C. II (2010). Clustering algorithms in biomedical research:a review. IEEE Rev. Biomed. Eng. 3, 120–154. doi: 10.1109/RBME.2010.2083647

Conflict of Interest: The authors declare that the research was conducted in theabsence of any commercial or financial relationships that could be construed as apotential conflict of interest.

Copyright © 2020 Winder, Sudduth, Fardo, Cheng, Goldstein, Nelson, Schmitt, JichaandWilcock. This is an open-access article distributed under the terms of the CreativeCommons Attribution License (CC BY). The use, distribution or reproduction inother forums is permitted, provided the original author(s) and the copyright owner(s)are credited and that the original publication in this journal is cited, in accordancewith accepted academic practice. No use, distribution or reproduction is permittedwhich does not comply with these terms.

Frontiers in Neuroscience | www.frontiersin.org 10 February 2020 | Volume 14 | Article 84